Tips and tricks
Habits that make Claude useful
None of these are complicated. Together they are the difference between fighting the tool and working with it. Try a few, keep the ones that fit how you work.
13 habits
Plan first
Ask for a plan first
For anything beyond a one-line request, ask Claude to plan before it acts, then read the plan. A plan is cheap to read and cheap to correct; undoing the wrong files is not. That is how a Cowork session is meant to run: you describe, Claude plans, you review, it executes, you steer.
Be specific
Be specific
Say the outcome you want, the constraints that matter, and where the files are. “Rewrite the about page so it is shorter and warmer, keep the headings, the file is in 03 web” gives Claude what it needs. “Make it better” does not.
Dictate
Dictate instead of typing
Speaking is faster than typing, and you give more detail when you talk. Press a hotkey, speak, and your words appear wherever your cursor is. Wispr Flow polishes as you go; Superwhisper runs entirely on your machine. Both have a free tier.
Memory
Use a CLAUDE.md
A short markdown file at the root of a project tells Claude who you are, how you want it to work, and your house rules, every time it opens that folder. It saves you re-explaining yourself. See The brain.
Tools
Let it use tools
Claude can search the web, look up documentation, and reach real services through MCP connectors. A lookup beats a guess. Connected tools are almost always better than the model working from memory alone.
A healthy task runs as a short loop
Context
Keep the context clean
Clutter makes Claude slower and less accurate. Use /clear to start fresh on an unrelated task, or /compact to keep going with a lighter history. For a big side-quest, ask Claude to use a subagent that runs one focused task and reports back.
Discover
Find skills with a slash
In Claude Code, type / (or /help) to see the skills and slash commands available, including ones from plugins and your own project skills. Learn /clear, /compact, /context and /usage first.
Effort
Match effort to the problem
Use a higher effort level for genuinely hard problems, and a faster mode for simple, low-stakes work; switch with /model and /config. Heavy settings and long sessions use more of your plan allowance, so save them for when they earn it.
Permissions
Set permissions deliberately
Pre-approve safe, repeated commands so Claude stops asking each time. Keep approvals tight for anything that changes or deletes data. For a connector or MCP tool, “Ask” is the safe default for anything that writes; reserve “Allow” for actions that only read.
Save it
Save what works
When Claude gets something right, lock it in. Ask it to remember a preference, or to write a reusable skill in your project’s .claude/skills/ folder. A preference in your CLAUDE.md reloads every session, so it survives /clear and /compact.
Evidence
Verify by looking
Do not trust “done” on its own. Ask Claude to run the thing and show you the output, or to take a screenshot of the web page, rather than reporting success without evidence. Seeing the result is the only reliable check.
Safety net
Use git
Keep your work in a git repository so every change is recorded and reversible. If an edit goes wrong, return to the last good state instead of trying to remember what changed. This single habit removes most of the fear from letting Claude edit files.
First move
Run a health check
When Claude Code is behaving oddly, run the built-in check. It inspects your installation and reports problems, the fastest first move.
claude doctor
Keep going
These tips, in the full guide, with everything around them.